1. Hybrid | 12-13 November 2020
Interciti Hotel, Daejeon, Republic of Korea
4th Conference on Nuclear Analytical Techniques (NAT2020)
Jointed with 6th Symposium on Radiation in Medicine, Space, and Power (RMSP-VI)
Deep-Learning based X-ray Image Classification for Quarantine Items:
a Feasibility Study
Yoonho Na1, Jimin Lee2 , Mingi Eom2 , Byung-Gun Park3 , Sung-Joon Ye1,4
1. Department of Applied Bioengineering, Graduate School of Convergence Science and Technology, Seoul National University, 08826, Seoul, Republic of Korea
2. Program in Biomedical Radiation Sciences, Department of Transdisciplinary Studies, Graduate School of Convergence Science and Technology, Seoul National University, 08826, Seoul, Republic of Korea
3. Neutron and Radioisotope Application Research Division, Korea Atomic Energy Research Institute, 34057, Daejeon, Republic of Korea
4.Advanced Institutes of Convergence Technology, Seoul National University, 16229, Suwon, Republic of Korea
(Corresponding author e-mail: sye@snu.ac.kr)
NAT2020-xxx
Introduction
As the time goes by, the number of tourist are increasing and accordingly, Illegal imports of quarantine items are rising.
Most of the quarantine of imported items is still carried out through detection dogs or simple X-ray visual monitoring.
Fast and efficient X-ray image identification system is needed instead of simple visual monitoring
Research Purpose : Investigate the applicability of the X-ray image recognition system using deep-learning for accurate X-ray
search and identification of quarantine items.
Experimental details / Methods Results
Discussion and Conclusions
References Acknowledgements / Funding support
Collecting X-ray Image data (courtesy of Byung-Gun Park)
The X-ray image has 21 classes with 13 different scan types.
21 classes : Pepper, Sesame, Dracaena, Lime, Banana, etc.
Preprocessing (courtesy of Mingi Eom & Jimin Lee)
Label the images with 21 classes and resize all the images
into same size.
Split data train : validation : test = 8 : 1 : 1
Model
The proposed model has 3 convolutional block, 1 flatten
layer and 2 fully connected layer.
The convolutional block is consisted with 3 3x3 filter
convolutional layer with leaky ReLU activation function.
Loss function : Cross entropy function
Optimizer : Adam optimizer
Epoch : 125
Learning rate : 0.0005, 0.0001, 0.00007, 0.00001
Number of trainable parameter : 106,942,182
This research was supported by a fund(Project Code No.PQ20205B030)
by Research of Animal and Plant Quarantine Agency, South Korea
Fig. 1. Deep-learning model architecture
Fig. 3. Confusion matrix from testset of quarantine object
Overall accuracy through classes : 25.5%
High accuracy : dracaena (67.57%), banana (42.67%)
Low accuracy : tree ear mushroom (8.33%), asparagus (8.57%)
Fig. 2. Single item and single item with carrier image of
Dracaena(top) and Tree ear mushroom(bottom)
Since the object visibility depends on the shape and density of the object, the classes like tree ear mushroom and asparagus had
bad visibility.
The lack of number of images per scan type lead the model hard to generalize the different object shapes and its labels.
Despite of the low overall accuracy, confusion matrix of the proposed model showed clear difference between each object, and
tendency for classifying objects.
The model have possibilities to be improved.